PhD Lecture Hall · August 26, 2026

How you ask changes the answer

One research result has stayed with me, because it proves that how we ask matters no less than the model itself.

Tôn Nữ Xuân Quyên August 26, 2026 2 min read
Card Day 22 — How you ask changes the answer
Card Day 22 — How you ask changes the answer

One research result has stayed with me, because it proves that how we ask matters no less than the model itself.

What prompt engineering is

Picture handing a hard problem to a bright member of staff. Ask straight out "what is the answer" and they are liable to reply in a hurry and get it wrong. Say instead "work through it step by step before concluding", and they slow down, reason in sequence, and land it far more often. Prompt engineering is the craft of writing the request so as to draw out exactly that way of thinking in AI, without touching the model at all.

A 2022 study by Wei and colleagues at Google showed that when the 540 billion parameter PaLM model was asked to lay out its reasoning step by step, an approach called chain-of-thought, on the GSM8K set of word problems, accuracy jumped from around 18% to nearly 57%. Same model, same problems, only a different way of setting the task.

How this applies to running a business

I find most disappointment with AI in business comes not from a weak model but from a task set too thinly. "Write me an advert" and "Write an advert for lipstick aimed at women aged 30 to 40, warm in tone, built around the feeling of confidence, with three headline options" produce results a long way apart.

Handing work to AI is exactly like handing it to a capable colleague who knows nothing about your context: the more clearly you state the role, the audience, the goal, the constraints and a worked example, the closer the result lands. This skill is cheap, quick to learn, and the thing a whole team should be fluent in before discussing any more complicated technology.

Looking back at the times AI disappointed you this past week, how much of it was a brief that was still vague?

To your desk

A clear brief usually produces a better result than a stronger model given a vague one.

This week you could try a small habit: before handing over a task, to a colleague or to AI, write three lines. One line of context, who this serves and what for. One line for the result you want, a concrete picture of the finished thing. And one line of worked example if you have one.

Those three lines take about two minutes, and usually save an afternoon of redoing. Setting work clearly from the start is one of the cheapest investments anyone running a business can make.

Bản tin

Mỗi ngày một thư ngắn.

Mỗi sáng tôi gửi một việc nhỏ bạn làm được ngay hôm đó với AI. Đọc hết trong hai phút.

Đăng ký xong bạn nhận luôn cẩm nang những câu lệnh tôi hay dùng nhất.
Bạn có thể huỷ bất cứ lúc nào.

Xong. Thư đầu tiên đang trên đường tới.